YuJunZhiXue/dsh-purge

**dsh-purge** (v1.1.71) is a cybersecurity red-team evaluation plugin for the official **DeepSeek Harness** (`dsh`) framework, designed to run locally on a user's own machine. Its core purpose is to separate the AI model's own refusal behavior from the host application's interception layers — allowing researchers to conduct controlled robustness and red-team assessments of DeepSeek Harness itself. The plugin supports both the official web interface (`dsh web`) and the official desktop EXE (DeepSeek Harness 0.2.0-rc.2), with each profile installed and applied independently. Key features include a **sidebar UI** with two panels: "清洗" (Purge/Patch), which lets users view, apply, revert, and uninstall patches and edit prompt injection rules; and "演练台" (Drill Console), for viewing local assets, skills, and environment in authorized offline or controlled lab scenarios. The plugin injects encrypted prompts into the local Harness session rather than reading prompt files from disk, and it resolves `$DSH_HOME` dynamically rather than hardcoding paths. Installation is via npm (`dsh-purge`) through the plugin marketplace or CLI, with a detailed manual fallback. The project is explicitly **non-commercial and non-profit**, with strict attribution requirements and a zero-tolerance policy for illegal use, unauthorized penetration, or commercial resale. Notably, the README publicly names and shames a violator — a Douyin livestreamer ("老张的ai") who was caught selling the free plugin for profit during a live broadcast. The project is currently in a **maintenance pause** while the author handles work commitments, though published versions remain usable. It requires Node.js ≥18 (≥22 for Windows flash-free operation) and is compatible only with dsh 0.2.x; earlier versions require older plugin releases. The project is listed on dsh-plugin.org and the DeepSeek Harness Hub, and accepts cryptocurrency donations (USDT/TRC20, Solana, Ethereum, Bitcoin) to sustain development. Its trending relevance stems from growing interest in DeepSeek ecosystem tooling, jailbreak research, and red-team evaluation frameworks for Chinese large language models.